MSRT: multi-scale representation transformer for regression-based human pose estimation

In this paper, we are interested in the human pose estimation problem with a focus on leveraging discriminative pose features. Recent pose estimation works concentrate on extracting high-level features but ignore the low-level details, thus reducing the prediction accuracy. To mitigate the above issues, we propose an end-to-end method called multi-scale representation transformer network (MSRT). Our network consists of two key components: feature aggregation module (FAM) and transformers. The FAM splits and stacks feature maps of different scales, then fuses them to achieve multi-scale representation learning. This module makes up for the lack of detailed information in the high-level features. Furthermore, we utilize Transformers to identify long-range interactions among feature maps, and capture implicit body structure information, which allows the proposed network to refine the locations of terminal and occluded joints. Compared with existing regression-based methods, MSRT achieves superior results on the COCO2017 and MPII datasets.